Choose optimization with measured evidence.

Inspect the hardest moments at full resolution

For each clip, the three seconds where the quality-target encode scored lowest, cropped at native pixels from the real encodes. Drag the divider: the H.264 baseline is on its left, quality-target HEVC on its right. Pause and step frame by frame. For browser playback the crops are re-encoded once to high-quality H.264, so treat the finest grain with care. In an evaluation you inspect the original files, on your own content.

Studio talk

H.264 BASELINEQUALITY-TARGET HEVC

Load comparison (1.9 MB)

Seconds 4.9–7.9 of the clip, a 960 × 540 crop at native pixels from a 1920 × 1080 frame. Mean VMAF over these three seconds: 94.79. Lowest frame: 92.88; the viewer opens on it.

Basketball

H.264 BASELINEQUALITY-TARGET HEVC

Load comparison (1.9 MB)

Seconds 23.4–26.4 of the clip, a 960 × 540 crop at native pixels from a 2048 × 1080 frame. Mean VMAF over these three seconds: 93.89. Lowest frame: 92.13; the viewer opens on it.

Sword action

H.264 BASELINEQUALITY-TARGET HEVC

Load comparison (5.1 MB)

Seconds 2.6–5.6 of the clip, a 960 × 540 crop at native pixels from a 1920 × 1080 frame. Mean VMAF over these three seconds: 95.64. Lowest frame: 90.58; the viewer opens on it.

Fireworks

H.264 BASELINEQUALITY-TARGET HEVC

Load comparison (9.8 MB)

Seconds 1.4–4.4 of the clip, a 960 × 540 crop at native pixels from a 2048 × 1080 frame. Mean VMAF over these three seconds: 95.23. Lowest frame: 91.02; the viewer opens on it.

Same codec: fixed-setting HEVC against quality-target HEVC

Same source, encoder, and preset (x265 slow, 60-frame lookahead); only the CRF differs. This isolates the effect of choosing the setting per title from the codec switch. It is still not an equal-quality comparison, because the fixed setting scores well above the target.

* Basketball figures cover the four 5-second samples used in the search (20 of 26.4 seconds). The other clips are complete.

What this result is, and is not

• Not an equal-quality comparison. The quality-target encodes score 0.3 to 2.5 VMAF points below the baseline. They were built to meet the target, not to match the baseline.
• Three effects are combined: switching from H.264 to HEVC, a slower encoder preset, and stopping at the quality target instead of exceeding it. HEVC needs compatible playback devices.
• Standard encoder controls. The encodes use x265. hls.ai selects the setting for each title and verifies the result; this is not proprietary compression.
• The floor is not a hard minimum. It bounds the 5th-percentile frame. The lowest single frame scored 89.3 in the fireworks clip.
• VMAF is an automated estimate. It does not replace visual review on your content and devices.
• Top rendition. These results are for the top rendition at native resolution.
• Sampling. A long title is searched on short samples rather than re-encoded in full. Here that applied to one 26-second clip. Compute cost per title-hour is measured on your content.
• Four stock clips. The setting search took 62–229 seconds per clip on one local machine, using five or six trial encodes.

Estimated reduction at the same VMAF

Relative to H.264 medium CRF 20 for each clip. Estimates interpolate log video bytes between measured quality points. These are not BD-rate values. Matching VMAF does not establish equal perceived quality; visual assessment is still needed.

The fireworks baseline VMAF is 99.84, near the score ceiling. Do not overinterpret small score differences or interpolation precision.

Processing cost belongs in the comparison

The experimental motion-aligned denoiser took about 71–73 seconds for each five-second excerpt and did not show a useful additional gain here. Preprocessing must demonstrate value beyond strong existing encoder baselines.

AV1 encoding took about 39–146 seconds per five-second excerpt. A local multithreaded encoder crash required single-thread libaom runs. These figures do not represent production product throughput.

Preprocessing research

An earlier experiment on the studio clip at 720p, with identical H.264 settings on both sides. Stable-background processing plus denoising reduced video data by 13%, with a lower VMAF score. Per-area frame rates made the file larger. This is separate from the quality-target results.

In this preprocessing experiment, VMAF is an automated quality estimate against the normalized source; higher is better. The combined treatment scored slightly below the baseline. Equal perceived quality has not yet been established.